By: Sylvain Collange (full.name.delete@this.gmail.com), August 13, 2011 7:54 am
Room: Moderated Discussions
ltcommander.data (ltcommander.tuvok@gmail.com) on 8/10/11 wrote:
---------------------------
>Is there something distinctively different between CS4.1 and OpenCL 1.0 that
>can't be worked around, which Intel seemed to have conveniently fallen into?
According to your link, Sandy Bridge does have some shared/local/group-shared memory, but it does not support random writes:
* Threads are limited to writing to only their own region of groupshared memory (although they can thread [read?] from any location).
* SV_GroupIndex must be used when accessing groupshared memory for writing.
This may explain the lack of OpenCL support.
AMD's RV670 has the same limitation, and is not OpenCL-compliant either, IIRC.
Numerical accuracy should not be an issue. DirectX 10.1 has mostly the same numerical hardware requirements as OpenCL 1.0.
---------------------------
>Is there something distinctively different between CS4.1 and OpenCL 1.0 that
>can't be worked around, which Intel seemed to have conveniently fallen into?
According to your link, Sandy Bridge does have some shared/local/group-shared memory, but it does not support random writes:
* Threads are limited to writing to only their own region of groupshared memory (although they can thread [read?] from any location).
* SV_GroupIndex must be used when accessing groupshared memory for writing.
This may explain the lack of OpenCL support.
AMD's RV670 has the same limitation, and is not OpenCL-compliant either, IIRC.
Numerical accuracy should not be an issue. DirectX 10.1 has mostly the same numerical hardware requirements as OpenCL 1.0.
